CrashFactory-VBD

VBD checkpoint for the behavior refinement of CrashFactory

📄 Paper | 🌐 Website | 💻 Code

VBD refines a reconstructed scenario into crash-consistent joint trajectories

vbd_default_e15.ckpt (137 MB) is the pretrained Versatile Behavior Diffusion (VBD) checkpoint that CrashFactory uses to refine reconstructed crashes into reactive, crash-consistent behavior. A scene encoder reads the road users, the map and the traffic lights, and a denoiser generates the joint trajectories of all road users while a guidance cost pulls the two crash participants to the reported collision point.

The checkpoint keeps VBD's architecture and was trained on the Waymo Open Motion Dataset, whose terms of use apply to the weights. Do not redistribute it.

🚦 Reactive Behavior

The same scenario with and without joint behavior modeling. With it, the left-turning vehicle (green) yields to the vehicle going straight (blue); without it, the vehicle follows its planned path and the two collide.

🚀 Usage

CrashFactory downloads the checkpoint on first use for accounts with access:

hf auth login
python scripts/run_behavior_refinement.py CrashResults/MTCF/crash_2021148118

A local copy can be passed with --vbd-checkpoint.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support